@nanmicoder/dsh-memory

English · 简体中文

dsh-memory · Long-term memory plugin for DeepSeek Harness

A two-phase memory pipeline (per-session extraction → global consolidation), a summary that is always injected into new sessions, and four on-demand retrieval tools.

Why long-term memory

Without memory, every session restates user preferences and re-steps known landmines. dsh-memory extracts durable knowledge from finished sessions and periodically consolidates it into a dense navigation summary (injected into every session) plus a grep-friendly handbook, so future sessions need fewer repeated instructions, waste fewer tool calls, and avoid known failure modes.

The design is adapted from the Codex memories system (two-phase extraction/consolidation, three-layer artifact layout, job claims, cooldown, redaction) and re-composed for DSH: no SQLite, no git baseline, no resident consolidation subagent.

Features

• 🧠 Phase 1 per-session extraction: after a session turns stop (debounced, 3 min default), its log is read, filtered, rendered, and redacted, then a model extracts structured memory (raw_memory + rollout_summary + slug). Low-signal sessions produce an empty no-op.
• 🧩 Phase 2 global consolidation: on a cooldown (6 h default) the new raw memories are merged into MEMORY.md (handbook) and memory_summary.md (dense index, exact v1 first line); raw input is rotated into an archive and never consolidated twice.
• 📥 Always-injected summary: memory_summary.md (hard size cap) is injected through the system prompt of every session — zero effort for the model to see the index.
• 🔍 Four memory tools: memory_list / memory_read / memory_search / memory_add (writes only on explicit user request).
• 🔒 Safety discipline: session content is treated as data, never instructions; secrets are redacted on both input and output; memory paths are confined to the memory root.
• 🔁 Reliable scheduling: one durable claim per session (KV-persisted); restarts never re-extract or double-consolidate; failures retry with backoff; orphaned claims recover.
• ⚙️ Settings page: Settings → Long-term memory for stats, manual runs, and config.

How it works

turn end ──► Phase 1 (per session) ──► rollout_summaries/.md + raw_memories.md
                                            │
                              (cooldown elapsed / new memories)
                                            ▼
                                     Phase 2 (global)
                                            │
              ┌─────────────────────────────┴────────────────────────────┐
              ▼                                                          ▼
   memory_summary.md (injected into every session)     MEMORY.md (searched on demand)